The Challenge of Subgroup Analyses
http://www.100md.com
《新英格兰医药杂志》
To the Editor: The Perspective article by Lagakos (April 20 issue)1 was a welcome explication of a contentious topic. Although the article focused on the role of chance and false positive results, it did not discuss another, more pernicious problem — bias. Whenever a subgroup analysis is performed, the randomization of patient characteristics between the treatment group and the control group is no longer necessarily maintained. Consider a subgroup analysis according to sex. The randomization process should ensure, if the sample is large enough, that the treatment and control groups are balanced according to sex. But randomization does not ensure that the two groups are balanced within the sex strata. If the men who received placebo are older and more severely ill than those in the treatment group, then the treatment may appear to be more beneficial among men, when in fact the result is due to the confounding effect of age and severity of illness. Specifying subgroups before the trial is conducted does not mitigate this bias; mitigation would require stratification according to the subgroup variable before randomization, so that patient characteristics would be balanced in the two groups within each subgroup stratum.
Mark D. Eisner, M.D., M.P.H.
University of California, San Francisco
San Francisco, CA 94117
mark.eisner@ucsf.edu
References
Lagakos SW. The challenge of subgroup analyses -- reporting without distorting. N Engl J Med 2006;354:1667-1669.
Dr. Lagakos replies: When performing ordinary randomization, we expect treatment groups to be balanced with respect to important patient characteristics, both in the entire sample and in any specific subgroup. However, randomization does not guarantee such balance, and when multiple subgroup analyses are conducted, the chances are increased that the treatment groups will be imbalanced with respect to important patient characteristics in at least one subgroup. Stratification of the randomization is one approach for ensuring balance with respect to certain factors, but the number of factors that can be controlled in this way is limited. Adjustment for patient characteristics in the analysis of the data is another, and such adjustment can therefore be useful in subgroup analyses. Finally, even when the treatment groups are balanced with respect to all important characteristics in every subgroup, false positive interactions between treatment group and a patient characteristic will still lead to biased (e.g., exaggerated) estimates of treatment differences within the corresponding subgroups, reinforcing the importance of properly controlling for type I errors in multiple-subgroup analyses.
Stephen W. Lagakos, Ph.D.
Harvard School of Public Health
Boston, MA 02115
Mark D. Eisner, M.D., M.P.H.
University of California, San Francisco
San Francisco, CA 94117
mark.eisner@ucsf.edu
References
Lagakos SW. The challenge of subgroup analyses -- reporting without distorting. N Engl J Med 2006;354:1667-1669.
Dr. Lagakos replies: When performing ordinary randomization, we expect treatment groups to be balanced with respect to important patient characteristics, both in the entire sample and in any specific subgroup. However, randomization does not guarantee such balance, and when multiple subgroup analyses are conducted, the chances are increased that the treatment groups will be imbalanced with respect to important patient characteristics in at least one subgroup. Stratification of the randomization is one approach for ensuring balance with respect to certain factors, but the number of factors that can be controlled in this way is limited. Adjustment for patient characteristics in the analysis of the data is another, and such adjustment can therefore be useful in subgroup analyses. Finally, even when the treatment groups are balanced with respect to all important characteristics in every subgroup, false positive interactions between treatment group and a patient characteristic will still lead to biased (e.g., exaggerated) estimates of treatment differences within the corresponding subgroups, reinforcing the importance of properly controlling for type I errors in multiple-subgroup analyses.
Stephen W. Lagakos, Ph.D.
Harvard School of Public Health
Boston, MA 02115